See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/267777956

Differentiating Ferrite and Martensite in Steel Microstructures Using Electron

Backscatter Diffraction

Article · January 2009

CITATIONS READS

22 9,527

3 authors, including:

Matt Nowell Stuart I Wright

EDAX EDAX
112 PUBLICATIONS 3,274 CITATIONS 212 PUBLICATIONS 6,425 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Matt Nowell on 26 March 2015.

The user has requested enhancement of the downloaded file.

 Differentiating Ferrite and Martensite in Steel Microstructures Using
Electron Backscatter Diffraction

M. M. Nowell, S. I. Wright and J. O. Carpenter

EDAX-TSL, Draper, Utah USA

Phase differentiation, martensite, ferrite, electron backscatter diffraction

Abstract

Electron Backscatter Diffraction (EBSD) has proven useful for characterizing the
crystallographic aspects of microstructure in steel. It is well suited to differentiating some phases
in steels such as austenite and ferrite. However, other constituent phases such as bainite,
martensite, cementite and pearlite are considerably more difficult. These difficulties arise
because of the similarity in crystallographic structure between constituent phases. For example,
martensite is body-centered tetragonal. Martensite typically has a c/a ratio of approximately 1.04
(the c/a ratio varies with carbon content). This is very close to a standard body-centered cubic
structure where the c/a ratio is exactly one. This similarity in crystal structure makes it difficult
to differentiate martensite from ferrite based using individual EBSD patterns. This work explores
different EBSD image quality based approaches for automated differentiation of martensite from
ferrite.

Introduction

Electron Backscatter Diffraction (EBSD) is an established characterization tool for

measuring point-specific crystallographic orientation with up to 10nm spatial resolution in
polycrystalline materials. As EBSD patterns are directly related to the crystallographic lattice
within the diffracting volume, EBSD is well suited to characterizing multiphase materials where
the crystal structures of the constituent phases are sufficiently distinct. For example, EBSD can
easily differentiate retained face-centered cubic (FCC) austenite from a body-centered cubic
(BCC) ferrite matrix. When the constituent phases are crystallographically similar but have
different chemical composition, the phases can be differentiated from each other using chemical
information collected simultaneously via Energy Dispersive Spectroscopy (EDS)1. However, in
steel the use of simultaneously collected EDS data to differentiate phases is very limited as the
differentiating element is typically carbon which is difficult to accurately characterize due to
both limitation in EDS measurements as well as the ubiquity of carbon in the environment.
While ferrite and austenite can be reliably differentiated using EBSD, differentiating
other constituent phases in steel microstructures can be quite challenging. Ferrite and austenite
are crystallographically distinct from each other, but other phases found in steel microstructures
are recognized by distinct microstructural characteristics as opposed to crystallographic
characteristics. For example, bainite is not an actual crystallographic phase but rather a specific
microstructure consisting of both ferrite and cementite. Thus, differentiating ferrite and bainite
from each other using individual EBSD patterns is impossible, as the unit cell is identical for
both bainite and ferrite.
 While many phases in steel could be considered, this work focuses on differentiating
ferrite from martensite. Martensite has a body centered tetragonal (BCT) structure. The c/a
lattice parameter ratio varies with carbon content but is typically near 1 and thus martensite can
be thought of as a pseudo-BCC structure. Martensite is, therefore, quite difficult to differentiate
from ferrite based by an analysis of the bands in EBSD patterns. However, EBSD patterns
obtained from martensite are generally of lower quality than patterns from ferrite. This is evident
in the patterns shown in Figure 1.

Figure 1 EBSD patterns from ferrite (a and c) and martensite (b and d) at 480 x 480 pixel resolution (a and b) and
96 x 96 pixel resolution (c and d).

Various metrics have been introduced to quantify the image quality (IQ) of diffraction
patterns. An IQ value based on the Hough transform2, 3 has been used to identify the area
fractions of ferrite and martensite by determining a threshold value in a bimodal IQ distribution 3.
The drawback of this approach is that a point just over the threshold value belonging to a
martensite grain would be associated incorrectly with a ferrite grain. This is particularly a
problem with points in the scan near grain boundaries where the IQ values are typically lower
than in the grain interiors. Some of these problems also exist for approaches based on fitting
multiple curves to the IQ distribution to measure multiple microstructural phases in steels4. This
method is purely statistical in nature and a one-to-one relationship between the individual points
in the scan and area fraction statistics cannot be established using such methods. This means that
subsequent orientation analysis of the constituent phases as determined by this methodology
cannot be achieved. The objective of this work is to improve not only area fraction
measurements of ferrite and martensite but also to enable each point in an OIM scan of dual
phase ferrite/martensite microstructure to be characterized as either ferrite or martensite in order
to accommodate subsequent analyses such as texture measurements from each phase or to
investigate orientation relationships across phases boundaries. It should be noted that image
quality is dependent on multiple variables5. IQ is a primarily a qualitative parameter making
comparative measurements on multiple specimens difficult. Even when stringently attempting to
hold all variables constant, absolute IQ values and distributions will change even when only
cycling the electron beam on and off. The analysis performed here will not focus on absolute IQ
values but relative values within individual scans.

Experimental Details

A dual phase low carbon steel specimen was mechanically polished6 for EBSD analysis.
EBSD data was obtained using EDAX-TSL OIM (version 5.3 alpha) systems on a thermal field
emission gun (FEG) FEI XL-30 scanning electron microscope (SEM) with a Hikari EBSD
detector and on a tungsten (W) source JEOL 6400 SEM with a DigiView 3 detector. EBSD
mapping data was collected from randomly selected 90µm x 90µm scan areas with a step size of
100nm for approximately 929,000 measurements per map. Each point was analyzed assuming a
BCC unit cell with a lattice parameter for ferrite (3.65Å). On the XL-30, a beam current of
approximately 6nA was used. Camera gain settings of 0%, 38%, and 80% (percent of maximum
gain) were selected, and camera exposure times were adjusted to optimize signal intensity.
Standard background subtraction was applied to each pattern during the scans except in the case
of one scan. In that case, an enhanced background correction routine using both static and
dynamic background correction and normalization of the intensity histogram was applied. This
was done simply for comparison to see the effects of additional image processing. As the FEG
SEM has a hard stage tilt stop at 75°, this tilt was used for all but one of the scans. In the other
case a tilt of 70° was selected for comparison. Similarly, the EBSD scans on the JEOL were
performed with a beam current of approximately 6nA and camera gain of 38% and a specimen
tilt of 70 degrees using a pre-tilted holder. For both detectors, EBSD patterns were analyzed at a
96 x 96 pixel resolution with a 12 bit dynamic range.

Results

Two different image quality metrics were calculated for each pattern in the scan: the
Hough transform image quality (IQH)2 and the intensity deviation (IQσ)7. A comparison of these
different metrics is detailed elsewhere5. Figure 1 shows IQH and IQσ maps, along with a map
constructed using a grain averaged IQH value (IQGA) for the points encompassed by each grain in
the map. A grain tolerance angle of 5° was used. The data used to construct the maps was
collected from the FEG with a 0% gain setting. The maps clearly show the dual phase nature of
the microstructure. The IQGA approach is included as it has been used to differentiate ferrite from
martensite previously8 (as well as deformed and recrystallized materials9).
Areas of ferrite and martensite were differentiated by manually selecting each ferrite
grain. A grain tolerance angle of 2° was used to define the grains; however; in some cases this
value was too large for accurate differentiation and values as low as 0.5° were used as discussed
in a subsequent section. The manually differentiated phase data is used as a “key” to analyze the
various IQ metrics. For each phase, the average (AVG) and standard deviation (SD) of each
image quality metric was calculated. It should be noted that points adjacent to grain boundaries
were excluded from the analysis4 to minimize any effects arising from the lower image quality
values typically found near grain boundaries. The phase contrast of each metric is defined by the
ratio of the ferrite average to the martensite average. A lower ratio signifies more contrast in the
IQ values between the phases which in turn leads to more robust differentiation. However, it is
emphasized that this ratio is not the sole metric of how well the two phases can be differentiated.
The width of each distribution is also important in quantifying how well the phases can be
distinguished from each other. The distribution overlap is defined as:

(1)

The smaller the overlap value, the less overlap between the two distributions. A negative
value indicates no overlap within two standard deviations. Two standard deviations were
selected because approximately 95% of the data is within this range of the mean for a normal
distribution. While the IQ distributions are not perfectly normal, they are approximated as such
for this analysis. Table 1 shows the ratio and overlap values for the IQH, IQσ, and IQGA
distributions under different acquisition conditions. Other IQ metrics7 were investigated as well.
However, these metrics produced contrast ratio values near unity and were thus excluded from
subsequent analysis. Figure 2 shows IQH, IQ and IQGA maps for the 0% gain FEG results.

Table 1 Ratio and overlap values for the different image quality metrics
Camera Gain Image IQH IQH IQσ IQσ IQGA IQGA
Source Tilt
(% Max) Processing Ratio Overlap Ratio Overlap Ratio Overlap
FEG 75° 0 subtraction 0.65 0.18 0.70 0.20 0.59 0.01
FEG 75° 38 subtraction 0.67 0.23 0.91 0.15 0.60 0.03
FEG 75° 80 subtraction 0.68 0.21 0.80 0.13 0.66 0.19
FEG 75° 80 enhanced 0.89 0.53 0.98 0.75 0.83 0.31
FEG 70° 80 subtraction 0.74 0.32 0.94 0.79 0.67 0.15
W 70° 38 subtraction 0.54 0.27 0.77 0.62 0.63 0.44

Figure 2 Scaled image quality maps for three different image quality metrics: (a) IQ H based on the Hough
transform, (b) IQ based on the average intensity deviation and (c) IQGA – grain averaged IQH values.

Discussion

Direct Analysis of Image Quality Results

In general, the IQGA values provide the best differentiation between the ferrite and
martensite phases. In every map collected using the FEG SEM, both the ratio and overlap values
are lower than the corresponding values for the IQH and IQσ distributions. The notable exception
to this trend is for the W results shows the highest IQGA value. It is also interesting to note that
the W data also produces the best IQ contrast ratio. This is due to the fact that for a given beam
current, the electron beam diameter of a W source SEM is significantly larger than that of a
thermal FEG SEM. Thus, the EBSD interaction volume is larger, and in the case of the
martensite phase, samples a higher fraction of both lower and higher angle grain boundaries.
This, in turn, also causes a higher standard deviation of the IQ values for both phases. For
example, the ratio of the standard deviation of IQGA to the average value of IQGA increases by a
factor of 2 for ferrite and 2.5 for the martensite. The overlap value is also increases. The effect
may be magnified by the lack of dynamic focus during mapping on this SEM. This capability
was only available for the FEG measurements. Without this capability, the electron beam was in
focus only along the horizontal center line of the map. As the beam is positioned away from the
horizontal center line, it becomes more defocused and the sampling volume is increased. This
observation suggests that defocusing the FEG beam may help improve contrast ratios but this has
yet to be explored.
The gain settings on the camera also affect on IQGA overlap values. The camera gain is
increased in order to reduce the exposure time necessary to obtain the optimal signal level.
However, increasing the gain also increases noise into the pattern. At the highest gain setting,
the increased noise caused a greater variation of measured IQ values within the grains, resulting
in a larger overlap value for IQGA. The best results were obtained with 0% gain. However, this
results in longer data collection times.
Image processing beyond the standard background subtraction, is often used in
multiphase measurements or when electron beam current stability issues are present to help
maintain a constant average EBSD pattern intensity. However, additional image processing was
deleterious to both the contrast ratios and overlaps. This is most likely due to the intensity
histogram normalization which essentially seeks to achieve an optimal contrast for each pattern.
Thus, the difference in pattern quality between the ferrite and martensite patterns is diminished.
It is therefore not recommended in this application. However, it may be advantageous in other
multiphase applications, particularly when the average atomic numbers of the constituent phases
differ significantly.
The tilt angle of the specimen also affects image quality. 70° tilt is often considered the
standard EBSD tilt condition. However, 70° is not a necessity, it evolved from using a piece of
(001) oriented single crystal silicon for pattern center calibration10. On EBSD systems, the EBSD
detector entered the chamber along a horizontal axis normal to the tilt axis of the system. In this
geometry, a (001) oriented single crystal tilted to 70° (or 69.6° more precisely) positions the
(112) zone axis at the location of the pattern center. This approach to pattern center calibration is
now rarely used as modern EBSD detectors enter the chamber at an angle from below so as to
optimize signal. EBSD patterns can be practically obtained at any tilt angle of 60° or greater.
While useable patterns can be obtained over a range of tilts, the IQ values change significantly
with tilt angle, increasing approximately 50% from 70° to 75° tilt, and approximately 100% from
70° to 80° tilt. In this work, it was found that increasing the tilt value improves the IQ contrast
ratios and overlaps. However these improvements were negated in the IQGA calculations.
Increasing the tilt increases the size of the interaction volume in the tilt direction. This
improvement in contrast ratios and overlaps with increasing interaction volume is in agreement
with the improved results obtained from the W measurements relative to the FEG results. This is
an area that may benefit from further investigation.
As noted previously, IQ values of scan points adjacent to grain boundaries are generally
lower than IQ values from points in the grain interiors. This is due to the beam interaction
volume sampling the grains on either side of the boundary as well as points along the boundary
plane itself. This leads to a composite pattern containing the superposition of the individual
patterns from both grains contained within the interaction volume. This not only lowers the IQ
value for the composite pattern but also makes reliable phase differentiation more difficult. In
addition, the fraction of points adjacent to boundaries directly depends on the step size of the
collection grid relative to the grain size of the material. The best FEG results were obtained using
0% gain at 75° tilt and standard background subtraction. This scan will be used in the following
more detailed analysis of the IQ data. From this scan, 22% of the measurement points lie next to
grain boundaries. The impact of including and excluding these boundary adjacent points is
shown in Table 2. When the boundary adjacent points are included, the contrast between phases
decreases, but the overlap for the IQH and IQσ values increases due to the increased spread in the
IQ values between the grain interiors and grain boundaries. However the IQGA spread decreases
into a negative value, indicating no overlap within two standard deviations. This is a result of the
coupling of the quality measurements with the orientation measurements used during grain
determination. This requires a high indexing success rate for both phases (99.6% in this case) as
using data cleanup procedures would likely arbitrarily mix the results between grains.

Table 2 Contrast ratios and overlaps including and excluding grain boundaries for the 0% gain FEG
measurements
Points adjacent to IQH IQH IQσ IQσ IQGA IQGA
grain boundaries Ratio Overlap Ratio Overlap Ratio Overlap
Excluded 0.65 0.18 0.70 0.20 0.59 0.01
Included 0.59 0.22 0.68 0.26 0.44 -0.07

Kernel Analysis of Image Quality Results

As noted in the previous discussion comparing the FEG and W results as well as the 70°
tilt and 75° tilt results, increasing the interaction volume seems to improve the results. This can
be approximated by smoothing the IQ data spatially within a map. One approach to accomplish
this smoothing, is to calculate the average IQ (IQKernel) within a kernel of a specified size. The
kernel, in this sense, is the set of points surrounding a given point in the scan out to the specified
number of neighbors. For the hexagonal scans used in this study, a first nearest neighbor kernel
would contain seven points – the point at the center and its six neighbors as shown in figure 3(a).

Figure 3 Schematic of (a) 1st, (b) 2nd, and (c) 3rd nearest neighbor kernels for a hexagonal grid.

The average IQ for the set of points in the kernel is then calculated and assigned to the
center point. The kernel averaging concept has been previously applied to local orientation
gradients for estimating local strain from EBSD measurements9. Maps constructed using this
kernel averaging approach are shown in figure 4. The smearing of the data clearly increases as
the kernel size increases.

Figure 4Kernel averaged image quality maps with kernel sizes of (a) 2nd, (b) 6th and (c) 10th nearest neighbors.
 Table 3 shows the average IQKernel values for the 2nd, 6th, 8th, and 10th nearest neighbors
including and excluding grain boundary adjacent points. When grain boundary adjacent points
are included, increasing the kernel size improves the overlap values, with the 10th neighbor
kernel producing an overlap value of -0.08, the best result achieved in this study. However, when
grain boundary adjacent points are excluded, an overlap minimum is reached at the 8th neighbor,
followed by an increase at the 10th neighbor. This suggests that both ferrite and martensite
measurements are beginning to be averaged together despite exclusion of the grain boundary
adjacent points at the larger kernel sizes which, in turn, increases the distribution overlap. The
IQKernel values also strongly depend on the kernel size relative to the step size.

Table 3 Contrast ratios and overlaps including and excluding grain boundaries for the 0% gain FEG
measurements
Points adjacent to IQKernel 2nd IQKernel 6th IQKernel 8th IQKernel 10th
grain boundaries Ratio Overlap Ratio Overlap Ratio Overlap Ratio Overlap
Excluded 0.65 0.11 0.65 0.00 0.65 -0.04 0.65 0.15
Included 0.59 0.14 0.59 0.00 0.59 -0.05 0.59 -0.08

Figure 5 shows phase maps for the manually differentiated phases, and for phases
differentiated using thresholding of the IQH (0.22 overlap), IQGA (-0.07 overlap), and IQ Kernel (-
0.08 overlap) values while including grain boundary adjacent points. Scan points associated with
ferrite are colored red and points associated with martensite are colored blue.

Figure 5 Phase maps (ferrite-red, martensite-blue, overlap-white) using a) manual differentiation and automated
thresholding based on b) IQH, c) IQGA, and d) 10th neighbor IQKernel. 5° boundaries shown in black
 In general, a scan point with an IQ value within two standard deviations of the average
IQ value of a given phase is assigned to that phase. Scan points with IQ values within the overlap
region are treated separately. For the IQH distribution, points within the overlapping region are
colored white. For the IQGA and IQKernel distributions where a negative overlap exists, a
numerical mean of the maximum value of the martensite distribution and the minimum value of
the ferrite distribution was used as a threshold to differentiate between the two phases. The
distributions and thresholding values corresponding to these maps are shown in figure 6.

Hough IQ Grain Ave IQ 10th Kernel IQ

0.08
0.07
0.06
Fraction

0.05
0.04
0.03
0.02
0.01
0.00
0 1000 2000 3000 4000
IQ
Figure 6 Image quality distributions color-coded for figure 5.

Differentiating using either the IQGA or IQKernel values provides improved results
compared to the IQH values. The IQKernel results most closely match the manual results. However
neither method is without some error. Figure 7 is a magnified view of the bottom right corner of
the map where the IQH, IQGA and IQKernel metrics all produced results different than those
obtained manually. The colored map is a grain map where each grain is colored uniquely relative
to its neighbors. A grain tolerance angle of 5°, the same as used for the IQGA calculations, was
selected for defining the neighbors. Two locations are labeled in the map shown in figure 7(b).
Location 1 shows a martensite grain where the average IQ is higher than the ferrite IQGA
threshold. Location 2 shows a junction between what appears to be a ferrite grain and a
martensite region. The misorientation between the martensite and ferrite is less than 2°. While
reducing the grain tolerance angle below this value separates these two grains specifically,
overall the reduction decreases the calculated grain size and reduces the effectiveness of the
grain averaged image quality values.
 Figure 7 (a) Image quality and (b) combined image quality and grain color map from the bottom left region of
figure 5 highlighting differentiation artifacts.

It was observed that the martensite grains are much more sensitive to the choice of grain
tolerance angle than the ferrite grains. Figure 8, shows a plot of the size of 3 selected martensite
grains and 3 selected ferrite grains as a function of the choice of grain tolerance angle. This plot
clearly shows that the ferrite grains maintain a nearly constant grain size regardless of the choice
of grain tolerance angle, while the martensite grains significantly change. This phenomenon may
also provide a means of differentiating ferrite and martensite in EBSD scan data.

Figure 8 Variation in area as a function of grain tolerance angle for three martensite grains and 3 ferrite grains

Another approach is local misorientation variation11 which in fact is related to the grain
size definition at smaller angles. Figure 9 shows the variation of orientation within 8 th nearest
neighbor kernels. Clearly, the martensite regions exhibit more orientation variation than the
ferrite grains. This approach is more quantitative than the IQ based approaches and may be more
transferrable from data set to data set. However, further analysis is needed to assess this
approach.
 Figure 9 Local orientation spreads in 10th nearest neighbor kernels.

Conclusion

Both the grain average image quality and kernel average image quality metrics provide
an improvement in accurately differentiating ferrite and martensite compared to methods based
on the image quality of individual scan points. The kernel method provides the best
differentiation and minimizes the number of errors relative to manual determination. It should
also be emphasize that careful selection of EBSD pattern acquisition parameters is also important
in achieving reliable results.

Acknowledgements

Thanks to J. Farrer of Brigham Young University for valuable discussions on the effect
of specimen tilt on EBSD image quality and A. Iza-Mendia of CEIT, San Sebastian, Spain for
providing the ferrite-martensite sample.

References

[1] M. M. Nowell and S. I. Wright, Phase differentiation via combined EBSD and XEDS.,
Journal of Microscopy Vol. 213, 2004, p 296-305

[2] K. Kunze, S. I. Wright, B. L. Adams and D. J. Dingley, Advances in Automatic EBSP Single
Orientation Measurements, Textures and Microstructures Vol. 20, 1993, p 41-54

[3] A. W. Wilson, J. D. Madison and G. Spanos, Determining phase volume fraction in steels by
electron backscattered diffraction, Scripta Materialia Vol. 45, 2001, p 1335-1340

[4] J. Wu, P. J. Wray, C. I. Garcia, M. Hua and A. J. DeArdo, Image quality analysis: A new
method of characterizing microstructures, ISIJ International Vol. 45, 2005, p 254-262
 [5] S. I. Wright and M. M. Nowell, EBSD Image Quality Mapping, Microscopy and
Microanalysis Vol. 12, 2006, p 72-84

[6] M. M. Nowell, R. A. Witt and B. W. True, EBSD Sample Preparation: Techniques, Tips and
Tricks, Microscopy Today Vol. 13, 2005, p 44-48

[7] X. Tao and A. Eades, Errors, Artifacts and Improvements in EBSD Processing and Mapping,
Microscopy and Microanalysis Vol. 11, 2005, p 79-87

[8] L. Ryde, Application of EBSD to analysis of microstructures in commercial steels, Materials

Science and Technology Vol. 22, 2006, p 1297-1306

[9] S. I. Wright, Quantification of Recrystallized Fraction from Orientation Imaging Scans,

Proceedings of the Twelfth International Conference on Textures of Materials (ICOTOM 12)
NRC Research Press, 1999, p 104-109

[10] D. J. Dingley and K. Baba-Kishi, Use of electron backscatter diffraction patterns for
determination of crystal symmetry elements, Scanning Electron Microscopy Vol. 2, 1986, p 383-
391

[11] S. Zaefferer, P. Romano and F. Friedel, EBSD as a tool to identify and quantify bainite and
ferrite in low-alloyed Al-TRIP steels, Journal of Microscopy Vol. 230, 2008, p 499-508

View publication stats